October 2026
The Legora BAR
The Benchmark for Agentic Reasoning measures how well AI models perform end-to-end legal tasks inside the Legora harness. Tasks are drawn from real-world cases, with outputs assessed against rubrics written by legal professionals. We repeat evaluations as both the models and the harness evolve. Each chart shows results relative to the average across the models evaluated for the relevant metric and task group.
October 2026 results
This month’s results cover 30 practice areas and seven task families across 17 models. Opus 5.5 leads on overall quality, scoring 6% above the 17-model average. It also leads in three of the seven task families, covering drafting, editing, and summarizing.
The models perform similarly on short tasks. The separation grows on longer cases, where the strongest model scores 55% higher than the weakest, compared with 7% on short tasks.
The overall ranking also hides real differences by type of work. Each of the six highest-ranked models drops to 10th or lower in at least one task family. It works the other way too: GPT-6.1 Sol, 13th overall, is in the top four on comparison tasks. Filtering by task family is the best way to choose a model for a specific job.
Quality by task length
Quality scores relative to the 17-model average for short, medium, and long tasks.
Quality vs. cost per case
Overall quality relative to the 17-model average, against average cost per case at public list prices.
Quality vs. median time per case
Overall quality relative to the 17-model average, against median minutes per case.
Citation grounding
Grounding score checks whether the source documents used to generate an answer are reflected in the citations.
Model verbosity
Average text output per case, split between responses and documents. Percentages show the difference from the 17-model average.
Quality by task family
Quality scores for each type of legal task, relative to the 17-model average for that task type.
Methodology
Each model completes the same legal tasks in isolated environments within the Legora platform, using the Legora agent harness. Each case is run three times, with outputs assessed by an LLM-as-a-judge against criteria written by legal professionals and weighted by importance. Benchmark cases use synthesized or anonymized data and remain private to reduce the risk of training contamination.
Built in collaboration with
